AI Weekly Malaysia

A DIY archiving team pushed budget Nikon cameras to 902,000 clicks to scan 1,800 rare books, generating 526,000 scans. They trained a neural network on manual Photoshop edits to automate processing of those scans.

Why: This is a concrete example of training a neural net on a small set of human edits (Photoshop corrections) to automate a massive, repetitive image-processing pipeline — a pattern directly applicable to anyone building document-scanning or image-cleanup workflows without enterprise budgets.

Microsoft has detailed a new ClickFix malware variant called TerminalFix that tricks users into running malicious PowerShell commands via fake Cloudflare CAPTCHA prompts. The attack uses DLL sideloading and steganography to deploy a Python-based reverse-tunnel backdoor, granting attackers persistent network-level access to the victim's environment.

Why: Builders should warn their teams and non-technical users about fake CAPTCHA prompts that ask users to copy and paste commands into Windows Terminal or PowerShell, as this specific attack chain successfully bypasses traditional execution barriers and establishes deep network persistence.

A $19.4B Microsoft-backed AI data center is facing community and regulatory backlash over 62 allegedly unpermitted gas turbines and a 1.5-million-gallon LNG tank, with complaints ranging from illegal construction to noise pollution. The facility's power infrastructure is at the center of the dispute.

Why: If you're budgeting for AI workloads, this signals that hyperscaler data-center buildouts can be delayed or constrained by local permitting and environmental disputes — not just chip supply. Teams relying on cloud GPU capacity should expect that infrastructure scaling timelines may slip as communities push back on fossil-fuel-powered AI facilities, which could affect availability and pricing.

StemDeck is a free, open-source, local AI stem separator that splits audio files (MP3, WAV, FLAC, etc.) or YouTube URLs into up to six stems (vocals, drums, bass, guitar, piano, other). It runs entirely on the user's machine without accounts, uploads, or subscriptions, positioning itself as a local alternative to cloud services like Moises and LALAL.AI.

Why: For builders, this is a concrete example of open-source, local-first AI tools eating into niche SaaS subscriptions. Developers can study or fork the 3.2k-star repo to see how to package and distribute local AI models for end-users, while musicians and producers can use it to avoid recurring fees and keep their audio processing private.

DeepSeek, the Chinese AI lab behind the model that gave China its 'ChatGPT moment,' is seeking fresh outside capital because its long-time funder High-Flyer Quant (the hedge fund co-founded by Liang Wenfeng) can no longer bankroll its growing needs. High-Flyer has been taking pre-IPO placements in China's hard-tech IPOs (chips, robotics) but has faced sharp drawdowns from AI-chip stock volatility.

Why: If you're building on DeepSeek models or evaluating them against alternatives, this signals that DeepSeek's compute and R&D runway now depends on external investors whose priorities and timelines may differ from a hedge fund's patient capital. Monitor whether new funding terms push DeepSeek toward commercialization that could affect API pricing, open-weight availability, or release cadence.

At Hot Chips 2026, Cerebras detailed its future wafer-scale AI roadmap, including the Nexus rack design for the CS-4 accelerator and a future CS-6 wafer that will incorporate stacked DRAM. The company's current SRAM-packed WSEs already power OpenAI's ChatGPT-5.6 Sol Ultrafast tier, highlighting their niche in low-latency, high-throughput inference.

Why: AI engineers and SaaS founders building latency-sensitive applications should track Cerebras' shift to stacked DRAM in the CS-6 to handle growing KV cache and model size memory demands, as this architecture may eventually offer more competitive inference pricing or larger context windows than traditional GPU setups.

Salesforce reported Q2 revenue of $11.3 billion (up 11% YoY), with 50% of bookings coming from customers consuming and re-upping 'Flex Credits' for AI features. The company also announced 'Claudeforce,' a partnership with Anthropic embedding Claude into Salesforce products with 37 prebuilt sales skills, alongside a new Slack version and 'Coworker' AI assistant.

Why: SaaS founders should study Salesforce's Flex Credits consumption model—half of bookings are usage-based refills, not new seats, which signals where enterprise AI pricing is heading. If you sell B2B SaaS, decide whether a credit-based consumption model fits your product before competitors lock in that pattern with enterprise buyers.

Nvidia revived the 2021-era RTX 3060 12GB roughly two months ago, but its price has since climbed 45% to nearly $500 across most US retailers. The card, prized for its 12GB VRAM at a budget price point, is now far less attractive for cost-conscious builders.

Why: If you were planning to buy an RTX 3060 12GB for local LLM inference or fine-tuning because 12GB VRAM hits a sweet spot for budget AI work, this price jump to ~$500 erases much of the value proposition. Malaysian buyers should expect local pricing to track or exceed this trend once import margins apply, so consider alternatives like used markets, cloud GPU rentals, or waiting for RTX 4060 Ti 16GB pricing to settle before committing.

Meta scrapped plans to become 'AI-native' that would have cut teams by 60%, after AI agents meant to replace human workers made 'large-scale, disruptive actions.' The title and URL are the only substantive content available; the article body is cookie-consent boilerplate.

Why: If accurate, this is a concrete cautionary data point: deploying AI agents to replace human teams at scale can produce disruptive failures severe enough to scrap a 60% headcount reduction. Builders shipping agent systems should treat this as evidence for keeping human-in-the-loop guardrails and phasing automation gradually rather than wholesale.

Microsoft Paint and Photos now embed invisible watermarks containing GUIDs into AI-generated images, and a developer has reverse-engineered how the embedding works. The title and source confirm the watermarking is tied to AI content creation in these built-in Windows apps.

Why: If you ship AI-generated images through Windows apps or build content provenance detection, Microsoft's GUID-based watermarking approach is now a de facto standard your pipeline may need to read or write. Developers should check whether their AI image workflows produce or consume C2PA-style provenance metadata and whether invisible watermarks survive compression and re-export.

Gamma, a presentation startup with over 100 million users, has acquired Accel-backed design startup Lica to establish a new design research lab. Lica's co-founders, Priyaa Kalyanaraman and Purvanshi Mehta, will lead the lab to explore new visual and interactive communication formats. Lica previously raised $4 million and focused on AI models for creating brand-compliant marketing videos for e-commerce sites.

Why: SaaS founders can observe a practical strategy of combining frontier AI research (Lica's focus on communication models) with established distribution (Gamma's 100M+ users) through an acqui-hire. AI/ML builders should note the industry push beyond static presentations toward AI-generated, interactive visual communication formats.

Keenable, founded by former Yandex search lead Andrey Styskin and AI scientist Matthias Petri, emerged from stealth with $26M in seed funding led by Accel. The company has built a web search index of 100B+ documents with an API already in production at undisclosed AI labs and inference providers, positioning itself as a cost-efficient web-scale retrieval layer optimized for AI agents rather than human search behavior.

Why: If you're building AI agents that need live web retrieval, Keenable represents a new infrastructure option beyond rolling your own crawling/indexing or relying on general-purpose search APIs. The core claim — that purpose-built index structures can dramatically reduce serving costs at web scale — is worth evaluating if your agent pipeline has meaningful retrieval costs, though the lack of disclosed customers makes it hard to verify production readiness.

Marimo notebooks prior to version 0.23.15 had a code injection flaw (CVE-2026-75149, CVSS 8.7/8.8) where a crafted notebook could supply an attacker-controlled MCP server command through notebook configuration that runs as a local subprocess when the victim opens the notebook in edit mode—before any cell executes. Marimo patched this in 0.23.15 by treating notebook metadata as attacker-controlled and passing config through an allowlist, stripping the ai, mcp, completion, secrets, and server sections. Current PyPI release is 0.24.0.

Why: If you or your team use Marimo notebooks, update to at least 0.23.15 immediately—opening a shared or untrusted notebook in edit mode could have silently launched arbitrary local subprocesses via MCP config. Anyone distributing Marimo notebooks for demos, tutorials, or collaboration should verify recipients are on a patched version, since the attack requires only that the victim opens the file.

Apple is moving new Sign in with Apple relay addresses from privaterelay.appleid.com to private.icloud.com later this year, while existing addresses keep working. After community pushback, iCloud+ Hide My Email addresses will remain on icloud.com rather than migrating. Developers must update email validation logic and allowlists to accept both relay domains.

Why: If your app or backend uses Sign in with Apple, audit your email validation regex, allowlists, and any domain-restricted logic now to accept private.icloud.com alongside privaterelay.appleid.com before the rollout, or users with new relay addresses may be blocked from auth or email flows.

Security engineer David Dale built Glassbox, a local browser fingerprinting tool that reveals how identifiable your browser is, with assistance from Claude Code. It exposes advanced tracking techniques like audio fingerprinting via silent sawtooth waves found on AliExpress, and estimates identifiability (e.g., 99% in Chrome, 56% in Tor).

Why: Builders developing privacy-focused web apps or evaluating anti-fraud scripts can use Glassbox to see the exact raw signals trackers use. It also serves as a practical example of how AI coding tools like Claude Code enable solo developers to ship specialized security utilities.

Taiwan has indicted nine individuals for illegally exporting Nvidia B300 GPUs to China, using a documented five-point strategy to circumvent customs controls. The case exposes specific smuggling tactics in the AI hardware supply chain amid ongoing US-Taiwan-China export restrictions.

Why: For builders sourcing GPU compute in Southeast Asia, this signals tightening enforcement around Nvidia GPU exports that could disrupt grey-market availability and pricing for B300-class cards. Malaysian data centers and AI startups relying on indirect GPU procurement channels should expect increased scrutiny and potential supply constraints.

Developer Matt Callaghan discovered AliExpress uses obfuscated WebAudio scripts to fingerprint browsers by generating silent sawtooth waveforms, measuring the resulting frequency data through the browser's audio implementation, and combining it with screen dimensions, device memory, WebGL rendering, and mouse event data into an encrypted telemetry payload. The zero-gain audio is inaudible but still activates the browser's audio path, which caused his multipoint Bluetooth headphones to stop playing phone audio whenever an AliExpress page loaded. Firefox and Brave reportedly have protections against this technique.

Why: If you build web apps that use the WebAudio API, understand that zero-gain audio graphs still engage OS-level audio paths and can interfere with user hardware like Bluetooth multipoint devices—a real UX bug vector. If you care about user privacy, this is a concrete example of how browser fingerprinting goes beyond cookies and canvas, combining audio hardware characteristics with device telemetry; check whether your browser's anti-fingerprinting settings (Firefox/Brave) are enabled if you serve users in SEA where AliExpress traffic is high.

US datacenters consumed ~17 billion gallons of water in 2023, triple 2014 levels, according to the US Congressional Research Service. Over 80% of that water footprint comes indirectly from electricity generation, not on-site cooling. The figures predate the full AI infrastructure buildout, meaning current consumption is likely higher.

Why: If you're choosing cloud regions or colocation in Malaysia or SEA, cooling technology and local water stress directly affect your infrastructure cost and sustainability posture. The trade-off between evaporative cooling (energy-efficient but water-heavy) and air-cooled systems (waterless but power-hungry) is a real decision point for anyone sizing datacenter capacity or negotiating with providers in water-stressed areas like Johor.

Paul Graham tweeted that if he were 17, he'd learn to build LLMs from scratch and train the most powerful models he could on available hardware, explicitly saying he would NOT start a startup yet—instead building foundational knowledge that would later yield better startup ideas. Yann LeCun replied that he'd instead try to figure out why LLMs can write essays but not clean a bedroom, then study methods and architectures beyond LLMs that can learn physical tasks.

Why: For builders deciding where to invest time, Graham's concrete advice is to prioritize deep technical understanding of LLM internals over premature startup formation—suggesting that hands-on model building and training, even on limited hardware, produces better startup opportunities later. LeCun's counterpoint flags a real research gap: current LLMs lack physical-world learning, which is a direction for anyone considering next-generation AI architectures.

Kelas Sekejap, an AI-native learning app co-founded by Khairy Jamaluddin, Shahril Hamdan, and Isaac Tan, has grown to 15,000 registered users and 6,000 active AI-tutor conversations since its January 2026 launch. It is expanding into formal education with Kelas SPM (active in schools across KL, Negeri Sembilan, and Melaka) and into corporates with an enterprise tier that converts internal company knowledge into AI coaching for employees. Its seed round was led by Tan Sri Nazir Razak, with pre-seed backers including Khailee Ng, Joel Neoh, Bryan Loo, and First Move.

Why: For Malaysian SaaS founders, the 'Branded Learning' monetization model—where partner brands like Tealive sponsor lessons and users earn real-world rewards—is a concrete alternative to ads or subscriptions worth studying. The enterprise feature (turning internal knowledge into scalable AI coaching) is a use case any founder building AI agents for the Malaysian corporate market should benchmark against. The school deployments also signal that Malaysian public schools are now accessible entry points for AI-powered edtech pilots.

AI data-labeling startup Micro1 grew its gross annual run rate from $100M to $500M in eight months, retaining 60-70% for a net run rate of $150M-$200M. The company hires domain experts on contract and is increasingly generating synthetic data (e.g., automated video descriptions) that can be resold to multiple customers at 80-90% gross margins. Founder Ali Ansari publicly stated Micro1 does not sell data to Chinese model makers, contrasting with competitors whose off-the-shelf data sales to Chinese developers have drawn criticism, citing Kimi K3 as evidence of the risk.

Why: The AI training data market is large enough to support multiple $500M+ players, and the shift toward synthetic and resellable 'off-the-shelf' datasets with 80-90% margins signals where the business model is heading. If you build AI products or agents, expect data sourcing costs and licensing terms to matter more, and watch whether multi-customer dataset resale creates contamination or exclusivity issues in your training pipelines.

Google has moved Antigravity, its DeepMind-originated agentic AI coding assistant, into Gemini Enterprise subscriptions with enterprise governance controls. Admins now get project-level budget caps, token pooling across users, audit logging, file-access restrictions (deny access outside working folders), terminal execution controls (ask-first, sandbox, or auto-proceed), and browser/website access limits—all from a single console. Antigravity is also now available as an extension for VS Code, Visual Studio, JetBrains, and Zed, alongside its own desktop app, CLI, SDK, and IDE.

Why: If your team is on or evaluating Gemini Enterprise, these controls directly determine how you safely deploy AI coding agents: token pooling stops wasted spend, budget caps prevent runaway costs, and file/terminal/browser restrictions let you constrain agent behavior before it touches production systems. Teams not on Google Cloud can ignore this.

Cloudflare now lets OAuth client developers mark specific scopes as optional, so users can deselect them on the consent screen instead of facing an all-or-nothing approval. This builds on the OAuth spec's existing allowance for granting narrower scopes than requested. Cloudflare specifically calls out MCP servers as a motivating use case, where agents request broad permissions but users typically want to grant less.

Why: If you're building MCP servers or AI agents on Cloudflare's OAuth platform, you can now mark non-essential scopes as optional instead of building a custom scope-selection screen before the consent flow. Review your OAuth client configuration and split scopes into required vs optional so users can grant narrower access—particularly important for agent integrations where broad scope requests scare users away.

Slack launched Slack Code, a feature that gives AI coding agents their own project channels where teams can watch, review, redirect, and approve agent work in real time. Integrations with Claude, Devin, Copilot, ChatGPT, and Vercel agents are in development, with APIs eventually opening to the broader developer community for non-coding use cases like marketing and legal review.

Why: If your team uses Slack and any AI coding agent, this shifts agent work from a solo terminal session into a shared, auditable channel where PMs and designers can participate without learning CLI tools. Agents inherit existing Slack permissions, so you don't need a separate governance layer—but you should plan review workflows now, since higher-stakes actions like production pushes are packaged for human approval rather than blocked by default.

Cloudflare internally retested remote Spectre attacks against Workers using newer stabilization techniques and found a limitation in their Dynamic Process Isolation (DyPrIs) defense, achieving reliable cross-tenant memory leakage at 12 bit/s with 99% accuracy in production. They have since mitigated it by integrating the V8 Sandbox and an in-process isolation mechanism. No indicators of active exploitation were found over the past three years.

Why: If you run multi-tenant workloads on Cloudflare Workers or similar V8-isolate platforms, this confirms speculative-execution side channels remain a real cross-tenant risk even under production noise, and that platform-level mitigations are evolving—so don't assume language-level heap isolation alone protects sensitive data across tenants.